Networks Ensemble for Multi-modal Cross-ethnicity Face Anti-spoofing
Cuiqun Chen, Meibin Qi · 2021
Face anti-spoofing is a crucial real-word detection problem in face recognition systems, which has a rapid development with the blooming of convolution neural network. The recently proposed CASIA-SURF CeFA dataset [1] is the largest multi-modal dataset containing 1607 subjects, 23583 videos and 2D plus 3D attack types, which is first time to consider the impact of cross-ethnicity for face anti-spoofing. In this paper, we adopt a simple yet effective multi-stream architecture which implements the feature extraction of each modality and the aggregation of the multi-modal features. Specifically, in order to alleviate the impact of the cross-ethnicity attacks and unknown spoofs proposed by Protocol 4 of the CASIA-SURF CeFA dataset, we design several networks trained separately based on the multi-stream architecture and then fuse the networks that with good robustness and generalization performance. In additional, we propose a data preprocessing method by enlarging the target area. The experimental results indicate the effectiveness of the proposed method.